Data security

[Freyvenn]

Hello everyone,

I have a question about Whisper, does it keep and store any data I feed it to train itself?

I think it should not as I am installing this locally, but in my case, I used google collab.

Just concerned about using it with confidential audio files as I work in law.

[jwnacnud]

If you are running it locally, you aren't sharing any data. If you use their API you can opt out of data sharing.

…

On Wed, Jun 21, 2023 at 10:53 AM Freyvenn ***@***.***> wrote: Hello everyone, I have a question about Whisper, does it keep and store any data I feed it to train itself? I think it should not as I am installing this locally, but in my case, I used google collab. Just concerned about using it with confidential audio files as I work in law. — Reply to this email directly, view it on GitHub <#1462>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAGW5A62SX2HAU5WCNIGUVTXMMRIDANCNFSM6AAAAAAZPA7CQ4> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[Freyvenn]

Is there an official post from openai stating that?

[ryanheise]

Is there an official post from openai stating that?

Not that I'm aware of, but it wouldn't really make a difference to how you can know whether it does or not. Even if OpenAI were to tell you one thing, the ultimate source of truth is the source code itself. And this is an Open Source project which means that nothing is hidden, you can see what code you are running, and you can see if it is opening any network connections to send or receive data.

In fact, if you look at the code, you can see that it is opening a network connection, but only to download data, not to send data. What data is it downloading? 1) a large file containing the neural net weights, and potentially 2) the audio file if you specify a URL.

I'm not saying you should believe me either, but the code is there. For example, take a look at the file init.py as it is the only file that imports the urllib module in order to make network connections, and it's doing so to download not upload. The other place is audio.py which downloads the audio file via ffmpeg. Poke around these files, as nothing is hidden, it's all laid bare.

(By the way, even if you are not able to understand the code, as others have pointed out elsewhere, you can also unplug your Internet connection before running Whisper and then you'd be fairly certain that it didn't send any data over the Internet.)

[nicholasgcotton]

I have a guide to getting Whisper set up to run completely offline here: https://github.com/nicholasgcotton/WhisperDO. IT's easiest to install it while online, but after a couple tweaks you can run Whisper without any network access at all. Why trust OpenAI when you don't have to?

[phineas-pta]

not everyone using windows

[nicholasgcotton]

I think most of the steps I outline can be made to work with Linux, you just have to figure out the .cache and vocab paths for yourself as I don't currently have a Linux box to test on.

[Freyvenn]

You're amazing, thank you for this!

…

________________________________ From: ryanheise ***@***.***> Sent: 22 June 2023 13:56 To: openai/whisper ***@***.***> Cc: Freyvenn ***@***.***>; Author ***@***.***> Subject: Re: [openai/whisper] Data security (Discussion #1462) Is there an official post from openai stating that? Not that I'm aware of, but it wouldn't really make a difference to how you can know whether it does or not. Even if OpenAI were to tell you one thing, the ultimate source of truth is the source code itself. And this is an Open Source project which means that nothing is hidden, you can see what code you are running, and you can see if it is opening any network connections to send or receive data. In fact, if you look at the code, you can see that it is opening a network connection, but only to download data, not to send data. What data is it downloading? 1) a large file containing the neural net weights, and potentially 2) the audio file if you specify a URL. I'm not saying you should believe me either, but the code is there. For example, take a look at the file init.py<https://github.com/openai/whisper/blob/main/whisper/__init__.py> as it is the only file that imports the urllib module in order to make network connections, and it's doing so to download not upload. The other place is audio.py<https://github.com/openai/whisper/blob/main/whisper/audio.py> which downloads the audio file via ffmpeg. Poke around these files, as nothing is hidden, it's all laid bare. (By the way, even if you are not able to understand the code, as others have pointed out elsewhere, you can also unplug your Internet connection before running Whisper and then you'd be fairly certain that it didn't send any data over the Internet.) — Reply to this email directly, view it on GitHub<#1462 (reply in thread)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/BAXVYJQWIYJB47RJKZP5BG3XMQ6IFANCNFSM6AAAAAAZPA7CQ4>. You are receiving this because you authored the thread.Message ID: ***@***.***>